Research Papers

Large Language Models (LLMs) tend to add disclaimers like "I'm just an AI" when asked about something related to themselves. The self-reports from such responses are used in debates about AI safety or self-knowledge of the models, yet what drives them is not well understood. Are the models telling us about themselves or rather how they are deployed? In this work, we show that the chat template works like a switch - when present, it turns this disclaimer voice up and experiential voice like "I feel" down, across 8 popular open-source instruct models up to 9B parameters in size. And conversely when the chat template is not present, it turns the disclaimer voice down and experiential voice up. Inside the activations of 3 models, we find a direction that steers this behavior. Removing the direction in the model's activation space turns disclaimer voice down and adding it turns it up, while a random direction of the same size has little effect. We find that instruct models without chat template, when we add the disclaimer direction to them, disclaim like the template was there. Since the chat template controls the disclaimer voice of LLMs, then researchers studying self-reports or introspection of models might have a confound they need to control for. Our results show that there is a direction they can use to steer this voice. More broadly, our work shows that what models say about themselves is not a fact about them. What they say doesn't come only from weights, but it is partially set by the chat template, and because of that a model's self-description shouldn't be treated literally.

The malicious use of generative artificial intelligence to create highly realistic deepfake videos raises serious ethical concerns and poses substantial challenges to AI safety. However, existing deepfake video benchmarks provide limited coverage of recent synthesis methods and generally lack reliable fine-grained textual annotations. Meanwhile, conventional detectors and multimodal large language models (MLLMs), whether operating as a single model or relying on a single analytical perspective, often fail to capture subtle forgery artifacts, limiting their generalization to emerging AI-generated methods. To address these limitations, we introduce FaceVid-Forensics-100K, a large-scale deepfake video dataset comprising 100,000 videos and spanning 33 synthesis methods across face swapping, face reenactment, and entire-face synthesis, including recent generators such as Seedance 2.0. The dataset provides fine-grained textual annotations of visual observations and verdict-consistent forensic explanations, automatically synthesized through a multi-model aggregation and conflict-resolution pipeline powered by advanced MLLMs. Building on this benchmark, we propose a multi-agent forensic reasoning framework that employs four specialized domain-expert agents to independently analyze forgery cues from four perspectives: texture, lighting, motion, and physics. A judge agent then reconciles their reports to produce a final prediction together with an explanation. Extensive evaluations on out-of-domain test sets show that, despite being composed entirely of small open-source MLLMs, our framework outperforms all methods including closed-source GPT and Gemini models and ranks first across all reported metrics on this benchmark. The project page is available at https://xavierjiezou.github.io/ARGUS/.

Edge Artificial Intelligence (Edge AI) enables the deployment of AI models directly on local edge devices, while such deployments are subject to strict resource constraints, particularly in clinical applications requiring local and timely inference. In such contexts, explainable artificial intelligence (XAI) can serve as a human-AI interface intended to support healthcare professionals' and patients' understanding of model predictions and informed decision-making. To fulfill this role, XAI method selection for TinyML deployments can be formulated as a human-centered multi-objective design problem that jointly considers qualitative stakeholder preferences, explanation quality, and proxy-based deployment cost. We propose a framework that integrates a large language model (LLM)-guided design interface that maps qualitative stakeholder preferences to candidate XAI methods, followed by deterministic feasibility filtering and Pareto-based optimization. The framework exposes trade-offs among explanation fidelity, stability, and proxy-based deployment cost while characterizing their implications for explanation quality and estimated deployment feasibility. A proof-of-concept evaluation on a skin lesion classification task illustrates how the framework systematically compares candidate XAI methods and identifies Pareto-efficient trade-offs. The present evaluation covers the computational selection stages, while physical MCU deployment and empirical human-expert validation remain outside the scope of this study.

As artificial intelligence (AI) systems transition from research prototypes to operational tools in Earth system science and forecasting, establishing trust in their predictions becomes increasingly important. Although model inputs and outputs are observable, the internal decision-making of modern AI models remains complex and hard to interpret, earning them the label ``black boxes.'' Explainable artificial intelligence (XAI) offers techniques to provide insight into these processes. However, most XAI methods were developed for classification tasks, raising questions about their suitability for the regression problems that dominate geoscientific applications. We review XAI approaches through this lens, organising them into a structured framework and examining both their theoretical foundations and practical behaviour. To ground this discussion, we apply a selection of methods to a machine learning emulator of the Lorenz 1963 system, an archetypal chaotic model that provides a tractable, physically meaningful setting for exposing the limitations and failure modes of general-purpose XAI in regression contexts. We then survey how these and related methods have been applied across a variety of Earth system sciences. We further situate XAI within the model development lifecycle, linking methodological choices to the needs of different stakeholder groups across operational Earth system science. We close by identifying gaps in existing methodologies and outlining a forward-looking research agenda, with practical recommendations for the responsible, effective use of XAI in regression applications of geoscientific modelling and forecasting.

Deep learning models have become integral to security applications due to their ability to model complex relationships in data and detect sophisticated threats. However, their complexity makes it difficult to understand how predictions are generated, posing significant challenges for interpretability, particularly in security applications where transparency is critical. Existing explanation methods, such as visual explanation techniques and post-hoc approaches, suffer from several limitations: reduced faithfulness due to local approximation errors, instability caused by reliance on randomness, and computational inefficiency that hinders real-time usage. To address these issues, we introduce XSec, a self-explainable deep architecture developed for security applications. During training, XSec uses a novel mask-based approach to extract informative sub-features from the data and learns prototypes, representative patterns that characterize each class. XSec then leverages the prototypes in a dedicated similarity layer at test time to compute similarity scores and generates interpretable explanations without the need for post-hoc analysis. We evaluate XSec across five diverse security scenarios, demonstrating its ability to achieve an average classification accuracy of 97.33% with minimal performance compromise. XSec produces deterministic explanations for a fixed trained model and input and substantially reduces explanation latency compared with approximation-based and perturbation-based post-hoc methods. Through this effort, we extend the applicability of self-explainable AI to security applications, bridging the gap between deep learning performance and the need for explainability in critical scenarios.

Artificial Intelligence (AI)-assisted navigation can help Arctic shipping adapt to rapidly changing sea-ice conditions, but reliable deployment requires reward models that are interpretable and robust to changing environments. Inverse reinforcement learning (IRL) provides a framework for recovering such rewards from vessel trajectories, while recent meta-IRL methods introduce latent context variables to capture behavioral heterogeneity. However, it remains unclear whether these latent representations recover genuinely hidden preferences or simply re-encode information already available in the observed state. We conduct a controlled evaluation on 3,186 AIS-derived voyages from 202 vessels across nine Arctic shipping seasons, comparing a linear shared reward, a nonlinear shared reward, and a latent-context model built on the same nonlinear architecture. The nonlinear reward improves held-out likelihood by 50.9% over the linear baseline, whereas adding vessel-specific latent context reduces performance by 16.5%. Behavioral analysis, context probes, and a pre-registered feature-hiding ablation show that apparent vessel-level variation is largely explained by observable route and environmental conditions rather than hidden vessel-specific factors. Moreover, predictive accuracy, route fidelity, and reward transfer yield different model rankings, demonstrating that no single metric is sufficient to evaluate learned rewards. These findings motivate testing whether the observed route, environmental, and vessel features already explain behavioral variation before adding per-vessel latent context. This supports more trustworthy AI deployment in safety-critical domains.

Challenges in Evaluating Explanation Methods for Static and Evolving Data

Jerzy Stefanowski

This paper addresses the limitations of Explainable Artificial Intelligence (XAI) with respect to insufficient evaluation. They are illustrated through the DetoxAI image recognition system for bias detection and concept unlearning. Then, an example of a human-grounded evaluation of methods for explaining image classification is presented. The paper further explores methods for adapting explanations to evolving data streams with concept drift. Experiences with adapting counterfactuals for this problem are discussed. Finally it is related to the challenges of tracking the co-evolution of data, models, and explanations.\footnote{This paper has been accepted for a publication in J.Nalepa (ed) Explainable AI in Space. Proceedings of EASi 2026 Workshop at IJCAI-ECAI 2026 Bremen, Springer CCIS vol 3107 (2016).}

Electronic health record (EHR) feature engineering is a major bottleneck in clinical research and AI, accounting for 39-45% of data scientists' workload. This is especially pronounced in heart failure, which affects an estimated 6.7 million U.S. adults and requires integrating fragmented EHR data with disease-specific, guideline-based clinical reasoning. Existing rule-based and large language model (LLM)-based approaches offer only partial automation with limited maintainability and evidence traceability. We developed the Nimblemind Multi-Agent System (nMAS), an evidence-linked, rubric-grounded pipeline for automated heart-failure feature engineering, and evaluated it on 500 dummy patient records from nine EHR source tables. nMAS generated 132 structured and 70 rubric-scored aggregated features, verified for structural integrity, rubric compliance, and provenance, and audited by a restricted LLM. Adding the aggregated features improved held-out AUROC from 0.895 to 0.963 for HFrEF and 0.870 to 0.910 for HFpEF phenotyping, and an independent LLM-based rubric assessment of evidence support and methodological soundness scored the features at 81.5% of maximum points. These results demonstrate the feasibility of automated, auditable feature engineering for complex cardiovascular EHR data, though evaluation was limited to a single-institution cohort and external validation is needed.

We introduce the neural echo as a tool for understanding the behavior of neural networks. It generalizes the model-based concepts of impulse responses, diffusion echoes, and filter echoes to learning-based methods. It provides local, space-adaptive impulse responses and filter kernels for a neural network, its so-called echoes. These echoes depend on the input image and can be visualized to understand the learned dynamics of the network via an affine mapping. Neural echoes build a bridge from classical signal processing to modern explainable AI. They are very general and can be applied to both image-to-image and classification networks, with convolutional or fully connected structure, of feedforward or recurrent type, including modern transformer networks. Network differentiability is not required. In the differentiable case, neural echoes comprise concepts based on the network Jacobian, such as saliency maps and the analysis of adversarial perturbations, as special instances. As a simple blueprint to explain our framework, we derive neural echoes for the denoising convolutional neural network (DnCNN). Our experiments suggest that this network weights pixels based on their spatial and gray value distances. This not only clarifies its behavior, but also shows that it can reproduce key concepts of classical model-based denoisers such as bilateral filtering.

Power indices, originating in cooperative game theory, quantify each player's influence on the outcome of a given game. Originally designed to distribute profits or costs among players and to analyse the fairness of voting systems, power indices have recently gained prominence as methods for attributing outputs of AI-based systems to inputs, thus facilitating explainability. However, selecting the appropriate power index for a given explanation task is an understudied problem. To address this, we organise power indices along three attribution dimensions: single-player, set-based, and cardinality-based. For each dimension, we review the corresponding power indices, generalise existing ones where applicable, and analyse which formal principles they satisfy. We provide proofs for properties that are missing in the literature and show that moving to the cardinality-based setting removes player-identity information while preserving some index-level distinctions. Using concrete examples, we illustrate how the choice of dimension and index affects the resulting attributions in practice, and offer guidance for practitioners seeking to select a suitable power index for a given application context.

Explaining the predictions of neural networks is a central challenge in trustworthy AI. Existing explanation methods, such as those based on feature attribution or minimal sufficient sets, typically treat input features as independent, which can yield misleading explanations when inputs exhibit structured dependencies. We address this by formalizing explanations as Halpern-Pearl (HP) actual causes, modeling input dependencies using Boolean Structural Causal Models (SCMs). We compute HP causes by applying bound propagation and branch-and-bound techniques, while providing formal guarantees of completeness and minimality. Our experiments show that we substantially outperform brute-force and ILP baselines in scalability, and outperform heuristic search as graph size grows, computing all minimal actual causes on instances with search spaces of up to $2.3\times10^{13}$ candidate (cause, contingency) pairs, on SCMs with up to 28 nodes, within a 180s per-instance budget. In a case study, we further show that ignoring input dependencies inflates the number of reported causes, 14.9% of which are spurious under our SCM.

Does Explainability Transfer? A Controlled Benchmark of Attribution Methods on Vision Transformers and CNNs

Sathiyamohan Nishankar, Nethmi Pathirana, Pubudu Sanjeewani +2

Most evidence on the effectiveness of explainable artificial intelligence (XAI) attribution methods has been established on convolutional neural networks (CNNs), with limited investigation into whether these conclusions generalize to the diverse Vision Transformer (ViT) architectures that now dominate computer vision. This paper presents a controlled benchmark that evaluates attribution quality across five dimensions: faithfulness, localization, robustness, complexity, and computational cost. A standardized framework assesses 13 attribution methods from four algorithmic families on eight representative backbones spanning CNNs, isotropic ViTs, hierarchical transformers, hybrid architectures, and linear-attention transformers. The results show that attribution performance is strongly architecture-dependent and that rankings established on CNNs do not reliably transfer to transformer-based models. CAM-based methods achieve the highest scores under the conventional bounding-box localization metric on CNNs and most ViTs but perform poorly on linear-attention architectures. Pixel-level dense-mask evaluation further reveals that these gains largely reflect metric saturation rather than accurate localization. CAM-based methods also exhibit limited robustness on global-attention transformers, whereas attention rollout provides consistently stable explanations with poor localization. Furthermore, faithfulness correlation offers limited discrimination between attribution methods, highlighting the limitations of single-metric evaluation. These findings challenge prevailing conclusions on attribution performance and demonstrate the need for architecture-aware, multi-dimensional evaluation. The open-source code for the evaluation framework and benchmark results is available at https://github.com/Nishan-Charlie/VIT_XAI_Bench.